6
$\begingroup$

What is the product of two independent student t distributions? In which case does this product product result in another t distribution?

$\endgroup$
5
  • $\begingroup$ Because exact formulas are messy and there is no product that results in a t distribution, would you mind explaining the statistical motivation for this question? We might be able to suggest alternative approaches or more useful related questions. $\endgroup$
    – whuber
    May 27 at 13:41
  • $\begingroup$ Explaining why I need this product would be a little bit complicated, but I am looking for a special case when the product of 2 t distribution will give a t distribution or an unnormalized t distribution $\endgroup$
    – sam
    May 27 at 15:18
  • $\begingroup$ Could you clarify what an "unnormalized" t distribution might be? $\endgroup$
    – whuber
    May 27 at 15:21
  • $\begingroup$ I reformulated my problem so it will result in answering the following question: stats.stackexchange.com/questions/576853/… $\endgroup$
    – sam
    May 27 at 16:37
  • $\begingroup$ The distribution corresponding to the product of several $t$ densities is called a poly-t distribution. $\endgroup$
    – Xi'an
    May 28 at 6:55

1 Answer 1

6
$\begingroup$

When $X$ and $Y$ are independent random variables with densities $f_X$ and $f_Y,$ the density of their product can be found with a change of variables as

$$f_{XY}(z) = \int_{\mathbb R} f_X(x) f_Y(z/x)\,\frac{\mathrm{d}x}{|x|}.$$

Ignoring normalizing constants (we'll consider these at the end), for two Student t densities with $\nu$ and $\mu$ degrees of freedom this integrand is proportional to

$$h(x,z) = \left(1 + \frac{x^2}{\mu}\right)^{-(\mu+1)/2}\, \left(1 + \frac{z^2}{x^2\nu}\right)^{-(\nu+1)/2}\,\frac{1}{|x|}.$$

Let's find a lower bound for $f(z)$ when $z$ is small. To do so, we may restrict the region of integration and we may replace the integrand by anything that never exceeds it.

Let $z$ be positive but less than $1$ and consider the integration region where $x^2\nu$ ranges between $z^2$ and $1.$ To get an appreciation for what's going on, here (with $\mu=\nu=1$) are plots of $h(x,z)$ for $|z| = 1$ (blue), $1/2, 1/4,$ and $1/8$ (red).

Figure

You can see that as $|z|$ approaches $0,$ there's more and more area pushed into this region. That's no surprise: we would expect the largest area (which corresponds to the highest density of the product) to be at the center of the product distribution, which (by symmetry) must be $0.$ But how large does it get?

Over the region $x^2/\nu \in[z^2, 1]$ the first factor of $h$ is smallest when $x$ is smallest and the second factor is smallest when $x$ is largest, whence throughout this region

$$\begin{aligned} h(x,z) &\ge \left(1 + \frac{z^2}{\mu\nu}\right)^{-(\mu+1)/2}\, \left(1 + \frac{z^2}{1}\right)^{-(\nu+1)/2}\,\frac{1}{|x|} \\ &\ge \left(1 + \frac{1}{\mu\nu}\right)^{-(\mu+1)/2}\, \left(1 + 1\right)^{-(\nu+1)/2}\,\frac{1}{|x|}. \end{aligned}$$

The second inequality is a consequence of $z^2 \le 1.$

The factors before $1/|x|$ are constant (but nonzero), depending only on $\mu$ and $\nu,$ so again let's consider them later and ignore them now. As $x$ varies over just the positive part of this region it runs from $z/\sqrt{\nu}$ to $1/\sqrt{\nu},$ giving a lower bound proportional to

$$\int_{z/\sqrt{\nu}}^{1/\sqrt{\nu}} \frac{\mathrm{d}x}{|x|} = -\log z.$$

As $z\to 0,$ this lower bound diverges. Consequently, no matter what the constants of proportionality are that we ignored, $f_{XY}(z)$ diverges at $0.$

Here, to illustrate, is a histogram from a simulation of ten million products (with $\nu=\mu=1/2$). Almost a million of those products are represented. The red curve is the negative logarithm. Clearly it approximates the density well near zero.

Figure

However, any Student t distribution with (say) $\kappa \gt 0$ degrees of freedom has a value proportional to $(1 + 0^2/\kappa)^{-(\kappa+1)/2} = 1$ at the origin, which is finite. Consequently, the product of two independent Student t distributions is never (even remotely like) a Student t distribution.


The product density can be found analytically as a polynomial combination of Riemann hypergeometric functions. Since this product is never a Student t distribution, though, I did not see any point into providing further details.

$\endgroup$
3
  • $\begingroup$ Thank you for this detailed analysis. this makes sense now. For the special case when the degree of freedom is 1 therefore the t distribution is a cauchy distribution, it is possible to find the result of the product of two t(1)? $\endgroup$
    – sam
    May 27 at 17:38
  • $\begingroup$ @sam It seems so, see here: math.stackexchange.com/a/1451744 $\endgroup$ May 27 at 17:47
  • $\begingroup$ Yes, that result is correct. But generally, for arbitrary degrees of freedom, the PDF of the product does not attain such a simple form: various versions of hypergeometric functions show up, such as elliptic functions. $\endgroup$
    – whuber
    May 27 at 18:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.